Automated Synthesis of Energy-Efficient Reconfigurable-Precision Circuits

Abstract

In recent years, designers are trying to move part of the computing tasks involved in Internet of Things applications from the cloud to the edge. This imposes increasing performance demands on edge nodes, which usually clash with their limited energy budget. An effective workaround is to leverage hardware capable of varying its computational precision at runtime, which can provide ''good-enough'' results while significantly reducing energy consumption. Dynamic Voltage and Accuracy Scaling (DVAS) and its variants are particularly promising methods to implement such hardware, due to their general applicability and contained overheads. However, these methods are negatively affected by the optimizations performed by commercial Electronic Design Automation (EDA) tools. As a consequence, when applied within a standard design flow, they do not yield the expected results. This paper describes a synthesis tool that solves this issue, allowing the integration of reconfigurable-precision circuits based on DVAS in standard design flows based on commercial tools. Moreover, our tool can receive information about the relevance of different precisions for the target application(s) that will use the circuit, and further optimize the design accordingly. When applied to two realistic use cases (neural network inference and image compression), our tool reduces the total energy consumption of reconfigurable-precision circuits of 20-25% compared to a straight-forward application of DVAS.